[vtk-developers] Speeding up raycasting on x86

[David Gobbi]

Hi Lisa,

I tried a little experiment with vtkVolumeRayCastCompositeFunction
and managed a speed-up of 35% in VTK's trilinear ray casting.

I replaced all instances of

 <something> = (int)(floatval)

with

 <something> = vtkQuickFloor(floatval)

according to the following definition of vtkQuickFloor:

inline int vtkQuickFloor(double x)
{
#if defined i386 || defined _M_IX86
  unsigned int hilo[2];
  *((double *)hilo) = x + 103079215104.0;  // (2**(52-16))*1.5
  return (int)((hilo[1]<<16)|(hilo[0]>>16));
#else
  return (int)(x);
#endif
}

That big mess with the bitshifts and whatnot is a trick that I found
on a web page a couple years ago (I could dig up the URL) that works
many times faster than what most iX86 compilers generate for (int)
casts.  I've been using it in vtkImageReslice for a while now.

Is this the sort of optimization that Kitware is interested in?
It is messy and CPU-specific, but it is confined to one single
inline function.

 - David